使用sklearn pairwise_distances计算X和y之间的距离相关

Question

I am currently trying various methods: 1. Correlation. 2. Mutual Information. 3. Distance Correlation to find the strength of relationship between the variables in X and the dependent variable in y. Correlation is the fastest and simplest(1 hour on a sample to 3 million records and 560 variables). Mutual Information calculation takes approximately 16 hours. I am also looking at distance correlation because of it's interesting property: The distance correlation between Xi and Y is zero , if and only if they are independent. However I am facing a problem while doing the calculation in Python.

below is my data:

X

prop_tenure prop_12m    prop_6m prop_3m 
0.04        0.04        0.06    0.08
0           0           0       0
0           0           0       0
0.06        0.06        0.1     0
0.38        0.38        0.25    0
0.61        0.61        0.66    0.61
0.01        0.01        0.02    0.02
0.1         0.1         0.12    0.16
0.04        0.04        0.04    0.09
0.22        0.22        0.22    0.22
0.72        0.72        0.73    0.72
0.39        0.39        0.45    0.64

**y**
status
0
0
1
1
0
0
0
1
0
0
0
1

I want to capture the distance correlation of each variable in X with y and store it in a dataframe and hence I am doing.

from sklearn.metrics.pairwise import pairwise_distances

num_metrics_df['distance_correlation'] = pairwise_distances(X,y,metric = 'correlation',njobs = -1)

However the documentation mentions the below:

If Y is given (default is None), then the returned matrix is the pairwise distance between the arrays from both X and Y.

This requires equal number of features in both X and Y?

How can I get distance correlation between each Xi and y in python? Can someone please help me with this?

Update:

I tried the approach of repeating the columns of y as per X.shape[1] and then do the calculation but it gives memory error for a sample of 10k records:

X = data_col.values

lb = preprocessing.LabelBinarizer()
df_target['drform'] = lb.fit_transform(df_target['status'])

y = df_target.values
n_rep = X.shape[1]
y = np.repeat(y,n_rep,axis = 1)

num_metrics_df['distance_correlation'] = pairwise_distances(X,y,metric = 'correlation',njobs = -1)

Traceback (most recent call last):

  File "<ipython-input-30-0f28f4b76a7e>", line 20, in <module>
    num_metrics_df['distance_correlation'] = pairwise_distances(X,y,metric = 'correlation',njobs = -1)

  File "C:\Users\test\AppData\Local\Continuum\anaconda3.1\lib\site-packages\sklearn\metrics\pairwise.py", line 1247, in pairwise_distances
    return _parallel_pairwise(X, Y, func, n_jobs, **kwds)

  File "C:\Users\test\AppData\Local\Continuum\anaconda3.1\lib\site-packages\sklearn\metrics\pairwise.py", line 1090, in _parallel_pairwise
    return func(X, Y, **kwds)

  File "C:\Users\test\AppData\Local\Continuum\anaconda3.1\lib\site-packages\scipy\spatial\distance.py", line 2381, in cdist
    dm = np.empty((mA, mB), dtype=np.double)

MemoryError

Answer 1

You can use scipy for this, although not explicitly parallelised, it is heavily optimised/vectorised. I find that it works super quickly for big datasets.

from scipy.spatial.distance import cdist
import numpy as np

n_samples = 100000
n_features = 50

X = np.random.random((n_samples, n_features))
y = np.random.choice([0, 1], size=(n_samples, 1))
correlations = cdist(X.T, y.T, metric='correlation')

But note that this returns a correlation distance, but there's a bunch of different metrics that you can use as well as custom metrics. More details are on the docs page .

Answer 2

Are you sure that you have computed what you wanted? It seems that scipy computes a distance based on (Pearson) correlation using this method. Maybe you wanted Székely's distance correlation, as in https://pypi.org/project/dcor/ .